This document indexes everything prepared for announcing and publishing FlashSpec, and gives you a clear sequence of actions.

Status: partially done.

✅ Initial measurements on Tesla T4 (Colab) with TinyLlama-1.1B-Chat NF4:

• 44.2 tok/s, α=0.75, p50=22.1ms
• Results: benchmarks/results/flashspec_ucb_tiny_llama.json
• Bandit regret figure: paper/figures/bandit_regret.jpg
• Gamma/draft-size sweeps: benchmarks/results/gamma_sweep.csv, draft_size_sweep.csv

⏳ Still needed for paper Table 1 and arXiv submission:

• H100 SXM5 runs with Llama-3-8B-Instruct, Llama-3-70B-Instruct, Mistral-7B
• Vanilla AR baseline on same hardware (to compute speedup ratio)
• Medusa and EAGLE baselines for the comparison table

export HF_TOKEN=hf_your_token
python scripts/download_models.py
make bench
git add benchmarks/results/ && git commit -m "bench: H100 results" && git push

This single action unblocks Zenodo:

1. GitHub repo → "Releases" → "Draft a new release"
2. Tag: v0.1.0
3. Title: FlashSpec v0.1.0 — Initial Release
4. Description: paste the relevant section from CHANGELOG.md
5. Publish

1. Go to https://zenodo.org, sign in with GitHub
2. Settings → GitHub → toggle on Mattral/FlashSpec
3. If you already created the release in Step 1, Zenodo auto-archives it and mints a DOI. If not, create the release now and it triggers automatically.
4. Copy the DOI badge Zenodo gives you
5. Add it to README.md (near the top, with the other badges) and to CITATION.cff under preferred-citation → add a doi: field

Why Zenodo first: it requires zero new writing, gives you a citable DOI within minutes, and the JOSS submission process explicitly checks for an archive link — having one ready makes JOSS review smoother.

JOSS's paper.md doesn't make performance claims, so it does not need to wait for benchmarks. Fill in the 3 placeholders in paper/joss/paper.md (ORCID, date, and reference the v0.1.0 tag from Step 1), then follow paper/joss/README.md to submit. Expect 2–8 weeks for review.

Once benchmarks/results/ has real numbers:

1. Fill in the abstract's [X.Xx] placeholder in paper/flashspec.tex
2. Fill in Table 1 with real numbers from benchmarks/results/*.json
3. Generate figures:

   jupyter nbconvert --to notebook --execute notebooks/02_bandit_analysis.ipynb
   jupyter nbconvert --to notebook --execute notebooks/03_kernel_profiling.ipynb

4. Compile: cd paper && make
5. Submit the resulting PDF + paper/ source to arxiv.org
   • Categories: cs.LG (primary) + cs.DC (cross-list)
   • Takes 1–2 business days for moderation

Fill in the real numbers in social/x_thread.md and social/linkedin_post.md. Best sequence:

1. Post once the arXiv ID exists — include the link in both posts
2. X first (faster-moving audience), LinkedIn within the same day
3. Reply to your own X thread with the LinkedIn link for cross-traffic

Preprint.org (ResearchGate's preprint server) has lower visibility in the ML/CS community than arXiv and is not indexed by Google Scholar as reliably. arXiv is the standard for ML systems papers and is what reviewers, conference PCs, and other researchers will look for. Use arXiv as primary; there's no need for a second preprint server once arXiv + Zenodo (for the software) + JOSS (for peer-reviewed software citation) are in place — between the three, you have priority, citability, and peer review covered.

Update this list as IDs become available:

• GitHub: https://github.com/Mattral/FlashSpec
• Docs: https://flashspec.readthedocs.io
• Zenodo DOI: TODO — paste after Step 2
• JOSS DOI: TODO — paste after Step 3 acceptance
• arXiv ID: TODO — paste after Step 4